Zichen Tian (Jason)
PhD candidate in Artificial Intelligence at Singapore Management University (SMU CVML Lab, advised by Prof. Qianru Sun). Previously research associate at S-Lab, Nanyang Technological University, and at Tsinghua University.
I study how foundation models learn where they can barely reason or verify, beyond the verification loop of recursive self-improvement. In this regime, observations are usually scarce and noisy. I reduce that noise until what they carry beyond the model's prior becomes verifiable, and then build it into, or extend, the prior's structure. This line of work includes debLoRA (NeurIPS 2024), MetaPEFT (CVPR 2025 Highlight) and mtLoRA (ICLR 2026). Methods: parameter-efficient fine-tuning (LoRA, PEFT), long-tailed and multi-task adaptation. Applications so far: satellite and radar imagery.
Honors
- Google PhD Fellowship 2026
- SMU Presidential Doctoral Fellowship 2026 and 2025
- SMU SCIS Dean's List 2026 and 2025
- CVPR 2025 Highlight (top 3% of submissions)
Selected Publications
- Scalable Multi-Task Low-Rank Model Adaptation (mtLoRA). Zichen Tian, Antoine Ledent, Qianru Sun. ICLR 2026. Scales multi-task LoRA to 15-25 tasks with Spectral-Aware Regularization, Block-Level Adaptation and Fine-Grained Routing.
- Meta-Learning Hyperparameters for Parameter Efficient Fine-Tuning (MetaPEFT / MetaLoRA). Zichen Tian, Yaoyao Liu, Qianru Sun. CVPR 2025 Highlight. Meta-learns PEFT module insertion, layer selection and module-wise learning rates for long-tailed remote-sensing adaptation.
- Learning De-Biased Representations for Remote-Sensing Imagery (debLoRA). Zichen Tian, Zhaozheng Chen, Qianru Sun. NeurIPS 2024. Unsupervised, LoRA-agnostic de-biasing of long-tailed remote-sensing features.
Profiles: ORCID · Google Scholar · Semantic Scholar · OpenAlex · DBLP · Hugging Face · ResearchGate · GitHub